AN

ALYST

FULL PA

PER

THE

www.rsc.org/analyst

Micro-Raman spectroscopy of natural and synthetic indigosamples

Peter Vandenabeele and Luc Moens

Ghent University, Laboratory of Analytical Chemistry, Proeftuinstraat 86, B-9000 Ghent,Belgium. E-mail: peter.vandenabeele@rug.ac.be; Tel: +32 9 264 66 23; Fax: +32 9 264 66 99

Received 1st October 2002, Accepted 10th January 2003First published as an Advance Article on the web 20th January 2003

In this work indigo samples from three different sources are studied by using Raman spectroscopy: the syntheticpigment and pigments from the woad (Isatis tinctoria) and the indigo plant (Indigofera tinctoria). 21 samples wereobtained from 8 suppliers; for each sample 5 Raman spectra were recorded and used for further chemometricalanalysis. Principal components analysis (PCA) was performed as data reduction method before applyinghierarchical cluster analysis. Linear discriminant analysis (LDA) was implemented as a non-hierarchicalsupervised pattern recognition method to build a classification model. In order to avoid broad-shaped interferencesfrom the fluorescence background, the influence of 1st and 2nd derivatives on the classification was studied byusing cross-validation. Although chemically identical, it is shown that Raman spectroscopy in combination withsuitable chemometric methods has the potential to discriminate between synthetic and natural indigo samples.

Introduction

For many years the blue colorant named indigo has been appliedas a textile dye as well as insoluble artists’ pigment.1 There areseveral natural sources of indigotin, the tropical genus In-digofera being the most well known, while in moderate climatezones (like western Europe) the woad plant (Isatis tinctoria)was used as an indigotin source. It lasted until 1880 when A.von Baeyer successfully synthesised synthetic indigo. K.Heumann published another synthesis route in 1890, whichformed the basis for the first commercial production by BASFin Ludwigshafen in 1897.2 It is clear that, when studying indigoapplied on textiles or on historical objects of art, positiveidentification of the synthetic product would enable to date theartefact post 1897. Unfortunately, synthetic and natural in-digotin (either from Indigofera or Isatis) are chemicallyidentical.

One theoretical way to distinguish between both types isbased on the 14C isotopic composition:3 the 14C/13C ratio insynthetic (i.e. from petrochemical origin) should be much lowerthan in indigo from more recent (natural) sources. However,experimental drawbacks associated with the relative largeamount of sample make this approach useless for art analysis.Indigoid colorants have been extensively studied by usingseveral spectroscopic techniques like micro-Raman spectros-copy,4,5 infrared-reflectance spectroscopy, electron spectro-scopic chemical analysis (ESCA) and proton-induced X-rayemission (PIXE).6 Thin-layer chromatography (TLC)7 has beenapplied to analyse indigo-containing samples, while it has beenshown that, by using high-performance liquid chromatography(HPLC), it was possible to distinguish indigo from woad andfrom Indigofera as the latter colorant contained small amountsof indirubin.8

Micro-Raman spectroscopy is appreciated as a non-destruc-tive method for art analysis that provides molecular informationon, among others, inorganic pigments,9–11 organic dyes,5organic binding media and varnishes,12 modern syntheticpigments,13 etc. Indigo has been identified by using thistechnique in contemporary textiles (jeans)14 and mediaevalmanuscripts.15 The colorant has been studied successfully byusing surface-enhanced resonance Raman spectroscopy

(SERRS)16 and band assignments, based on ab initio calcula-tions, have been proposed.17

This work focuses on the ability of micro-Raman spectros-copy in combination with several chemometric methods todistinguish between natural and synthetic indigo samples.Principal components analysis (PCA) and spectral pre-process-ing methods were applied to extract relevant information fromthe molecular Raman spectra. By performing cluster analysisand linear discriminant analysis (LDA) it is shown which latentvariables contain relevant information for the discrimination.

Theory

Several chemometric methods are applied in order to differ-entiate between indigo samples of different origin. Principalcomponents analysis is a data extraction method that reducesthe number of variables to a limited number of latent variablesthat contain maximal variance. Cluster analysis is an un-supervised classification method, while linear discriminantanalysis (LDA) is a supervised pattern recognition method (i.e.a training set with spectra of well-known nature (whethersynthetic or natural) is involved).

Principal components analysis (PCA)

Principal components analysis is a data reduction method that isfrequently used for the analysis of spectroscopic data.18–20 Eachof the n spectra, all containing m variables, can be considered asa point in an m-dimensional space. During PCA, a new set ofaxes is constructed such that each principal component axis isorthogonal to the others and captures maximal variance over thespectra: every spectrum can be described by its coordinates(called scores) on these principal components axes.21,22 Thescore sip on the pth principal components axis for each spectrumi can be considered as a linear combination of the intensities (I)for all m original band positions:

(1)

This journal is © The Royal Society of Chemistry 2003

DOI: 10.1039/b209630g Analyst, 2003, 128, 187–193 187

Publ

ishe

d on

20

Janu

ary

2003

. Dow

nloa

ded

by N

orth

east

ern

Uni

vers

ity o

n 26

/10/

2014

23:

19:5

5.

View Article Online / Journal Homepage / Table of Contents for this issue

The loading njp reflects the weight of the original band positionj in the calculation of the pth principal component. It is foundthat the dataset can be adequately described in a k-dimensionalprincipal components space, the number of dimensions (k)generally being much smaller than the original number ofvariables (m). The m–k remaining variables are consideredredundant or noise. The variables Iij may be original intensities,but they may as well be the result of some pre-processing.

Cluster analysis

By performing hierarchical cluster analysis,21 cases (i.e.spectra) are grouped according to their similarity. Differentclustering algorithms and (dis)similarity measures are availa-ble.23 In this work simple Euclidean distance Dhi is used, inorder to measure the (dis)similarity between two spectra h andi:

(2)

Cluster analysis is not always performed on the originalvariables I, but latent variables, such as the scores in the k-dimensional principal components space are frequently used:

(2a)

The clustering method that is applied in this work is the averagelinkage method: an object is clustered with a group of objects onbasis of the average Euclidian distance to all the objects in thegroup.

Linear discriminant analysis (LDA)

Linear discriminant analysis is a supervised classificationmethod, which is used for the discrimination of samples in apreviously well-known number of classes.21,22 The dataset issplit into a training set and a validation set: the training set isused for model building, while the validation set is used forevaluation of the model. LDA is a linear method that minimises

the within-group variance and maximises the between-groupvariance. During LDA, for each spectrum i, k new variables si

are defined, dubbed canonical variates, which are a linearcombination of the original intensities Ii of the spectrum.

(3)

The factor wjp indicates the weight of the jth variable on thecalculation of the pth LDA axis; the weights are found bycalculating the eigen vectors of the matrix-equation:

VMw = lw (4)

The matrix M describes the distance between the means of thedifferent classes (i.e. the between-group distance, which ismaximised by LDA) while the V matrix can be considered as ameasure for the spread within the different groups. Opposite toPCA, in LDA these latent variables are selected not to capturemaximal variance over all the cases, however to capturemaximal separation over the classes. The cases are plotted in thethus-defined LDA-space and classification happens accordingto the shortest distance towards the centroids of the differentgroups in the training set. One of the requirements for LDA isthat the number of cases (n) has to be smaller than the numberof variables (m) and therefore, in spectroscopic applications apreceding data-reduction is often performed (e.g. PCA).

Experimental

21 powdered samples of indigo were obtained from 8 differentsuppliers worldwide. An overview is given in Table 1. Sinceoutliers may strongly affect the chemometrical analysis, thesample S-7 was not used for these investigations, as the supplierstipulated that this sample contained only 60% of indigo. 6 ofthe remaining 20 samples were of synthetic origin, 14 of themoriginated from natural sources (11 Indigofera and 3 Isatis). Forevery sample, 5 Raman spectra were recorded of differentpigment grains, in order to capture accidental within-samplevariance.

Raman analysis was performed by using a Renishaw System-1000 instrument, equipped with a 785 nm diode laser with apower output of 50 mW. In order to avoid crystalline and

Table 1 Overview of the origin of the indigo samples that were analysed during this work. Because of the impurities, sample S-7 was not used for thechemometrical analysis

Sample codea Supplierb Provenancec Remarksc

I-1 George Weil (Germany)I-2 Weaving Southwest (USA) IndiaI-3 Livos (Germany) India Dating 1995I-4 Livos (Germany) India Dating 1996I-5 Livos (Germany) India Dating 1992I-6 Livos (Germany) India Dating 1998I-7 Livos (Germany) India Dating 1995I-8 Maiwa (Canada) Azules—El SalvadorI-9 Maiwa (Canada) Miani Sinah—PakistanI-10 Maiwa (Canada) Pitchie Reddy Vallur AP—IndiaI-11 Galke (Germany) MilledW-1 Livos (Germany) Germany Dating 1995W-2 Livos (Germany) United Kingdom Dating 2001W-3 Livos (Germany) United Kingdom Dating 2001S-1 Kremer (Germany)S-2 George Weil (Germany)S-3 Weaving Southwest (USA) USAS-4 Schmincke (Germany)S-5 Chemische Fabriek Triade B.V. (The Netherlands) China Pure indigo powderS-6 Chemische Fabriek Triade B.V. (The Netherlands) Russia Pure indigo powderS-7 Chemische Fabriek Triade B.V. (The Netherlands) Germany 60% Purea Sample Code: the letter indicates the origin of the indigo sample: S: synthetic production; I: indigo plant (Indigofera); W: woad plant (Isatis). b Supplier:the addresses of the different suppliers can be obtained from the authors on request. c Provenance/Remarks: according to the supplier.

188 Analyst, 2003, 128, 187–193

Publ

ishe

d on

20

Janu

ary

2003

. Dow

nloa

ded

by N

orth

east

ern

Uni

vers

ity o

n 26

/10/

2014

23:

19:5

5.

View Article Online

isomerical transformations, the laser power was further reducedwith a neutral density filter with an optical throughput of 1%.Spectra were recorded by using a 503 enlargement objective,enabling the recording of spectra of distinct particles with adiameter of ca. 2 mm. By using a 1200 lines/mm grating and aPeltier-cooled CCD detector, a spectral resolution of ca. 1 cm21

could be achieved. Raman spectra were accumulated over 300 sin the spectral range from 300 till 1800 cm21.

Raman spectra, recorded by using Wire software (Renishaw,Wotton-under-Edge) working under Grams (Galactic indus-tries), were exported to ASCII format and subsequentlyanalysed by using Matlab 6 (Eigenvector Research, Inc.), thePLS toolbox 2.1 (Eigenvector Research, Inc.) and the Dis-criminant Analysis Toolbox 0.3 (Michael Kiefte, University ofAlberta, Canada).

Results and discussion

The Raman spectrum of indigo

For every sample 5 spectra from different pigment grains wererecorded. The similarity between a natural and a syntheticindigo sample is shown in Fig. 1, next to a baseline correctedspectrum. Indigotin in the trans-configuration is a flat cen-trosymmetric molecule and thus, IR-active vibrations are notdetected by Raman spectroscopy and vice versa.

In general, the band positions correspond well with theextensive spectral interpretation by Tatsch and Schrader,17

which is based on NIR-FT-Raman spectra (l0 = 1064 nm) andab initio calculations. On the other hand, for some bands,different band intensities are noticed: for instance, the verystrong band at 544 cm21 has been reported as being of mediumintensity. It is very likely that these observations are caused byresonance enhancement of some vibrations by the 785 nm laser.In the Raman spectra of our indigo samples, at some positions,doublets (1631/1623 cm21 and 868/859 cm21) or shoulders(1606 cm21, 1541 cm21, 775 cm21) are observed, althoughthey were not reported by Tatsch and Schrader. Different

crystalline structures, eventually caused by inclusions of otherisomers, may be the origin of these shoulders. For instance, it isnot unlikely that some cis- or leuco-indigotin molecules disturbthe crystalline structure and thus influence the recorded Ramanspectrum.

Principal components analysis and cluster analysis

The spectra of our indigo samples are very similar and simpleobservation makes it is hard to tell whether the spectrumcorresponds with a natural or a synthetic sample. In order toreduce the number of variables and to be able to distinguishbetween relevant and accidental fluctuations, principal compo-nents analysis (PCA) is used as a data extraction method. Beforeperforming PCA, all the spectra were vector-normalised, as,

Fig. 2 Dendrogram resulting from the hierarchical cluster analysis of 100 Raman spectra of 20 indigo samples. Clustering was performed by using anaverage linkage algorithm and Euclidean distance on the first 3 principal components. Synthetic indigo samples are shaded

Fig. 1 Raman spectra of a natural (I-2, a) and synthetic (S-3, b) indigosample. For clarity a baseline corrected spectrum (subtraction of a 6thdegree polynomial) of b was added. Raman spectra were recorded by usinga 785 nm laser (0.5 mW at the source) and an accumulation time of 300 s.

Analyst, 2003, 128, 187–193 189

Publ

ishe

d on

20

Janu

ary

2003

. Dow

nloa

ded

by N

orth

east

ern

Uni

vers

ity o

n 26

/10/

2014

23:

19:5

5.

View Article Online

among others, slight changes in measurement geometry over thedifferent spectra may result in different over-all intensities.From every band position Ij in the spectrum the averageintensity µ over all band positions in the spectrum is subtractedand then divided by the standard deviation s of the peakintensities:

(5)

By performing principal components analysis of these samples,99.9% of the variance was captured in the first 3 PCs, whichwere retained for the further investigations.

Hierarchical cluster analysis was applied on 100 cases byusing average linkage and Euclidian distance; the dendrogram isshown in Fig. 2. As PCA transforms the original variables intoorthogonal latent variables, the use of Mahalanobis distance isnot required in order to take correlation into account. Whencomparing this result, obtained with unscaled data, with thoseobtained by using previously scaled data (mean centred orautoscaled), the latter classifications were worse and theclusters tended to form chains of cases, introducing moremisclassifications.

In this dendrogram all but two of the spectra of the syntheticindigo samples cluster together. When studying these outliers,spectra with a differently shaped fluorescence backgroundoccur, which hamper the distinction. It should be remarked thatfor both of the cases, the four other spectra taken from the samesample cluster together with synthetic indigo. The distinctionbetween Indigofera and Isatis cannot be made: woad samples

Fig. 3 Scree plot and score plots corresponding with the loading plots in Fig. 4 and the dendrogram in Fig. 2. The labels i, w and s correspond to indigo,woad and synthetic indigo respectively.

Fig. 4 Loading plots for the first 3 principal components that are extractedfrom 100 vector-normalised spectra from 20 indigo samples.

190 Analyst, 2003, 128, 187–193

Publ

ishe

d on

20

Janu

ary

2003

. Dow

nloa

ded

by N

orth

east

ern

Uni

vers

ity o

n 26

/10/

2014

23:

19:5

5.

View Article Online

are spread over the natural indigo branch. Spectra taken fromdifferent pigment grains from the same sample are notnecessary tightly connected, which indicates that there is aconsiderable within-sample variance.

During this classification experiment, 3 principal compo-nents were used. The scree plot, as well as the score plots arepresented in Fig. 3. The score plots give a graphical representa-tion of the spread of the different samples in the PC space.Loading plots express the contribution of the different variablesin the calculation of the respective PCs. In these loading plots(Fig. 4), though more pronounced in those of PC 2 and PC 3, thespectral features from the indigo spectrum can easily berecognised. Moreover, the loading plots reveal that there is asignificant contribution of the background in the calculation ofthe principal components. In Raman spectroscopy, the spectral

Table 2 Summary of the results from the linear discriminant analysis ofthe indigo samples (n = 100). Numbers are the percentage correct classifiedspectra, according to the cross validation as stipulated in the text.Parameters are the pre-processing techniques, number of classes todiscriminate (2: Natural/Synthetic; 3: Indigofera/Isatis/Synthetic) anddifferent principal components that are used for the analysis

Pre-processing

Numberof classes

PCs 1 to2

PCs 1 to3

PCs 1 to4

PCs 2 to3

PCs 2to 4

No derivative 2 98 98 98 98 981st derivative 2 92 91 93 90 952nd derivative 2 71 93 100 97 100No derivative 3 83 87 82 86 841st derivative 3 86 86 84 82 852nd derivative 3 66 88 90 93 90

Fig. 5 Score plots for the first 4 PCs, corresponding with the chemometrical analysis in Fig. 4. The marks correspond to synthetic indigo (s) and the naturalgenera Indigofera (i) and Isatis (w).

Analyst, 2003, 128, 187–193 191

Publ

ishe

d on

20

Janu

ary

2003

. Dow

nloa

ded

by N

orth

east

ern

Uni

vers

ity o

n 26

/10/

2014

23:

19:5

5.

View Article Online

background often originates from broader spectral features likefluorescence or phosphorescence. These may be caused by theanalyte as well by impurities or by the matrix and they mayoverwhelm the component-specific Raman spectrum.

As it is aimed to obtain a wide applicable classificationalgorithm, it is desirable to reduce the contribution of the

fluorescence background. Spectral pre-processing techniques,such as derivatives or second derivatives, can be applied inorder to eliminate the contribution of broad features (comparedto Raman bands) before autoscaling and principal componentanalysis is performed.

Until now in this work, hierarchical cluster analysis was usedas a classification method. This approach has the advantage thatthe relationship between the different samples is representedeasily in a dendrogram. However, this approach is hardly opento objectification as there is no measure for a ‘good’classification. Moreover, if unknown samples have to becharacterised, the whole analysis has to be performed again, andhopefully the unknown spectrum clusters tight with one oranother spectrum in the database. Each new spectrum affects theclustering tree and may disturb any previously describedclassification rule. Therefore another classification method wasselected, namely linear discriminant analysis.

Pre-processing and linear discriminant analysis: crossvalidation

The aim of this work is to obtain the best possible discriminationbetween natural and synthetic indigo samples. Therefore,different pre-processing techniques are combined with theextraction of various sets of PCs, before performing LDA. Allthe spectra have been examined, once as such, after taking thefirst and second derivatives (Savitsky-Golay algorithm, windowwidth: 21 datapoints and 2nd degree polynomial).24 Subse-quently, all the spectra were autoscaled (eqn. (5)), in order toeliminate multiplicative effects. In the next step, principalcomponents analysis was performed and a specified number ofPCs was extracted for the linear discriminant analysis. LDA isa supervised classification technique, requiring a training set ofspectra and a validation set. Therefore, the dataset was split into3 parts, taking care that all the spectra of a particular samplebelong to the same group and that the spectra of each type ofsamples (synthetic, indigo, woad) were evenly distributed overthe 3 groups. The groups contained 30, 35 and 35 spectra each.For each analysis, LDA was performed 3 times, each time thetraining set being compiled of two groups, the remaining groupbeing the training set. The number of correct classified casesover the 3 analyses was added, and is a measure for the quality

Fig. 6 Loading plots for the first 4 principal components that are extractedfrom 100 vector-normalised 2nd derivatives of spectra from 20 indigosamples.

Fig. 7 Details from the Raman spectra of natural and synthetic indigo samples in the spectral regions between 300 and 500 cm21 and between 1500 and1650 cm21. Raman spectra were recorded by using a 785 nm laser (0.5 mW at the source) and an accumulation time of 300 s.

192 Analyst, 2003, 128, 187–193

Publ

ishe

d on

20

Janu

ary

2003

. Dow

nloa

ded

by N

orth

east

ern

Uni

vers

ity o

n 26

/10/

2014

23:

19:5

5.

View Article Online

of the LDA model. The results of the investigations aresummarized in Table 2.

When comparing the upper half of the table with the lowerpart, it is easily seen that discrimination between two classesyields better classification results than classification into 3groups. This can be partially explained from the a prioriprobabilities, but when studying different biplots of PCs againsteach other, it is seen that, although small, the group of Isatisspectra contains a large variance. Therefore, we focus on thedistinction between natural and synthetic indigo.

The constant results for the classification of the unprocessedindigo spectra are remarkable. From the loading plots (Fig. 4),it is clear that the contribution from the background isconsiderable in all the PCs and may dominate the classificationresults. Therefore, 1st and 2nd derivatives were taken from allthe spectra. From Table 2 it can be seen that, independent ofwhich PCs are selected, the 1st derivative spectra yield worseresults than the untreated or 2nd derivative spectra. Whencomparing the last four columns, it is seen that, thoughincluding the 1st PC provides a broader basis for the LDA, theresults are often worse. This means that the 1st principalcomponent is less significant for the discrimination between theindigo samples. This is as well illustrated in the score plots (Fig.5). When performing the calculations by using only the first twoPCs, the models obtained yield low percentages correctclassified spectra: it seems preferable to select a datasetconsisting of 3 or 4 PCs. By comparing the results from thecross-validation and bearing in mind that the first PC tends todestabilise the model, LDA based on PCs 2 to 4 yields the bestclassification results.

Loading plots for the first 4 principal components of the 2ndderivative spectra are shown in Fig. 6. Opposite to the loadingsof the unprocessed spectra (Fig. 4), there is no contribution fromthe background, providing a model that is less sensitive tofluorescence. The general appearance of these loading plots ismuch more complex than the first ones, as the shape of thesecond derivative of a single Raman band is even more complexthan the shape of the band itself: whereas a single Lorentzianhas one maximum, its 2nd derivative has 2 maxima and aminimum. It is clear from these loading plots that the Ramanbands at 1582 and 1571 cm21 contribute to the discriminationbetween natural and synthetic indigo. These bands can beassociated with symmetric (Ag) stretching vibrations ofv(C = C), v(C = O) and v(C–C). From the spectra (Fig. 7) itappears that these spectral features have a different bandwidth:these bands are better resolved in the synthetic indigo spectra.The spectral region between 300 and 900 cm21 in the loadingplots of PCs 3 and 4 appears noisy, but nevertheless thecontribution seems to be significant to the classification. Whencomparing these loadings with the spectral features, it turns outthat this is not an artefact, as the loadings correspond to weakspectral features, e.g. at ca. 310, 400 and 460 cm21 (Fig. 7).Spectral features below 1000 cm21 can be attributed to ringvibrations and d(C–H), d(NCC), d(C = O) and d(C = C–C = O).By taking the 2nd derivative, their presence is emphasisedleading to a significant contribution to the discriminationbetween natural and synthetic indigo samples.

Conclusions

In this work Raman spectroscopy has been applied to studysynthetic and natural (Indigofera or Isatis) indigo samples ofdifferent provenance. By using several chemometrics tech-niques, it was possible to distinguish between these groups.Hierarchical cluster analysis was performed as an unsupervisedclassification method and yielded good results. The dataset was

reduced by using principal components before applying lineardiscriminant analysis on the spectra and on their 1st and 2ndderivative. Cross-validation showed that the best discriminationwas obtained by using 2nd derivatives and up to 4 PCs. Bycomparing the loading plots, it was possible to identify thespectral features that sign for the discrimination between naturaland synthetic indigo species.

Acknowledgements

The authors are in debt to all people and companies that assistedin this research. They are especially grateful to the suppliers ofthe indigo samples: George Weil, Weaving Southwest, Livos,Maiwa, Galke, Schmicke and Chemische Fabriek Triade B.V.The authors wish to thank the research council of the GhentUniversity (B.O.F.) and the Fund for Scientific Research—Flanders (F.W.O.-Vlaanderen) for their financial support. P.V.is especially grateful for his postdoctoral fellowship (Post-doctoraal onderzoeker) from F.W.O.-Vlaanderen.

References

1 H. Schweppe, Handbuch der Naturfarbstoffe—Vorkommen, Verwen-dung, Nachweis, Ecomed Velagsgesellschaft, Landsberg/Lech,1993.

2 E. Steingruber, in UllmannAs Encyclopedia of Industrial Chemistry:Indigo and Indigoid Colorants, ed. B. Elvers, J. F. Rounsaville and G.Schulz, VCH Verlagsgesellschaft, Weinheim, 1990.

3 R. J. H. Clark, C. J. Cooksey, M. A. M. Daniels and R. Withnall,Endeavour, New Series, 1993, 17, 191.

4 R. Withnall, A. Derbyshire, S. Thiel and M. J. Hughes, Proc. SPIE-Int. Soc. Opt. Eng, 2000, 4098, 217.

5 G. N. Andreev, B. Schrader, H. Schulz, R. Fuchs, S. Popov and N.Handjieva, Fresenius’ J. Anal. Chem., 2001, 371, 1009.

6 P. E. McGovern and R. H. Michel, Anal. Chem., 1985, 57, 1515A.7 P. E. McGovern and R. H. Michel, Acc. Chem. Res., 1990, 23, 152.8 J. Wouters, Dyes in History and Art, 1992, 10, 17.9 H. G. M. Edwards, D. W. Farwell, E. N. Newton and F. Rull Perez,

Analyst, 1999, 124, 1323.10 P. Vandenabeele, B. Wehling, L. Moens, B. Dekeyzer, B. Cardon, A.

von Bohlen and R. Klockenkämpe, Analyst, 1999, 124, 169.11 D. C. Smith, M. Bouchard and M. Lorblanchet, J. Raman Spectrosc.,

1999, 30, 347.12 P. Vandenabeele, B. Wehling, L. Moens, H. Edwards, M. De Reu and

G. Van Hooydonk, Anal. Chim. Acta, 2000, 407, 261.13 P. Vandenabeele, L. Moens, H. G. M. Edwards and R. Dams, J.

Raman Spectrosc., 2000, 31, 509.14 C. Coupry, G. Sagon and P. Gorguet-Ballesteros, J. Raman

Spectrosc., 1997, 28, 85.15 S. P. Best, R. J. H. Clark and R. Withnall, Endeavour, New Series,

1992, 16, 66.16 I. T. Shadi, B. Z. Chowdhry and R. Withnall, Conference Proceed-

ings of the 9th ECSBM—European Conference on the Spectroscopyof Biological Molecules, Vol. B, Prague, 2001, p B226.

17 E. Tatsch and B. Schrader, J. Raman Spectrosc., 1995, 26, 467.18 M. Shimoyama, H. Maeda, I. Sato, T. Ninomiya and Y. Ozaki, Appl.

Spectrosc., 1997, 51, 1154.19 P. Vandenabeele, A. Hardy, H. G. M. Edwards and L. Moens, Appl.

Spectrosc., 2001, 55, 525.20 R. H. Brody, H. G. M. Edwards and A. M. Pollard, Anal. Chim. Acta,

2001, 427, 223.21 D. L. Massart, B. G. M. Vandeginste, L. M. C. Buydens, S. De Jong,

P. J. Lewi and J. Smeyers-Verbeke, Handbook of Chemometrics andQualimetrics, Elsevier, Amsterdam, 1997.

22 M. Otto, Chemometrics—Statistics and Computer Application inAnalytical Chemistry, Wiley-VCH, Weinheim, 1999.

23 J. A. Hartigan, Clustering Algorithms, John Wiley and Sons, NewYork, 1975.

24 A. Savitzky and M. J. E. Golay, Anal. Chem., 1964, 36, 1627.

Analyst, 2003, 128, 187–193 193

Publ

ishe

d on

20

Janu

ary

2003

. Dow

nloa

ded

by N

orth

east

ern

Uni

vers

ity o

n 26

/10/

2014

23:

19:5

5.

View Article Online